SpinNet: Learning a General Surface Descriptor for 3D Point Cloud Registration (CVPR 2021)

This is the official repository of SpinNet, a conceptually simple neural architecture to extract local
features which are rotationally invariant whilst sufficiently informative to enable accurate registration. For technical details, please refer to:

SpinNet: Learning a General Surface Descriptor for 3D Point Cloud Registration

Sheng Ao*, Qingyong Hu*, Bo Yang, Andrew Markham, Yulan Guo.

( indicates equal contribution*)

[Paper] [Video] [Project page]

(1) Overview

(2) Setup

This code has been tested with Python 3.6, Pytorch 1.6.0, CUDA 10.2 on Ubuntu 18.04.

Clone the repository

git clone https://github.com/QingyongHu/SpinNet && cd SpinNet

Setup conda virtual environment

conda create -n spinnet python=3.6
source activate spinnet
conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.2 -c pytorch
conda install -c open3d-admin open3d==0.11.1
pip install "git+git://github.com/erikwijmans/Pointnet2_PyTorch.git#egg=pointnet2_ops&subdirectory=pointnet2_ops_lib"

(3) 3DMatch

Download the processed dataset from Google Drive, Baidu Yun (Verification code:d1vn) and put the folder into data.
Then the structure should be as follows:

--data--3DMatch--fragments
              |--intermediate-files-real
              |--patches

Training

Training SpinNet on the 3DMatch dataset:

cd ./ThreeDMatch/Train
python train.py

Testing

Evaluate the performance of the trained models on the 3DMatch dataset:

cd ./ThreeDMatch/Test
python preparation.py

The learned descriptors for each point will be saved in ThreeDMatch/Test/SpinNet_{timestr}/ folder.
Then the Feature Matching Recall(FMR) and Inlier Ratio(IR) can be calculated by running:

python evaluate.py [timestr]

The ground truth poses have been put in the ThreeDMatch/Test/gt_result folder.
The Registration Recall can be calculated by running the evaluate.m in ThreeDMatch/Test/3dmatch which are provided by 3DMatch.
Note that, you need to modify the descriptorName to SpinNet_{timestr} in the ThreeDMatch/Test/3dmatch/evaluate.m file.

(4) KITTI

Download the processed dataset from Google Drive, Baidu Yun (Verification code:d1vn), and put the folder into data.
Then the structure is as follows:

--data--KITTI--dataset
            |--icp
            |--patches

Training

Training SpinNet on the KITTI dataset:

cd ./KITTI/Train/
python train.py

Testing

Evaluate the performance of the trained models on the KITTI dataset:

cd ./KITTI/Test/
python test_kitti.py

(5) ETH

The test set can be downloaded from here, and put the folder into data, then the structure is as follows:

--data--ETH--gazebo_summer
          |--gazebo_winter
          |--wood_autmn
          |--wood_summer

(6) Generalization across Unseen Datasets

3DMatch to ETH

Generalization from 3DMatch dataset to ETH dataset:

cd ./generalization/ThreeDMatch-to-ETH
python preparation.py

The descriptors for each point will be generated and saved in the generalization/ThreeDMatch-to-ETH/SpinNet_{timestr}/ folder.
Then the Feature Matching Recall and inlier ratio can be caluclated by running

python evaluate.py [timestr]

3DMatch to KITTI

Generalization from 3DMatch dataset to KITTI dataset:

cd ./generalization/ThreeDMatch-to-KITTI
python test.py

KITTI to 3DMatch

Generalization from KITTI dataset to 3DMatch dataset:

cd ./generalization/KITTI-to-ThreeDMatch
python preparation.py

The descriptors for each point will be generated and saved in generalization/KITTI-to-3DMatch/SpinNet_{timestr}/ folder.
Then the Feature Matching Recall and inlier ratio can be caluclated by running

python evaluate.py [timestr]

Acknowledgement

In this project, we use (parts of) the implementations of the following works:

Citation

If you find our work useful in your research, please consider citing:

@inproceedings{ao2020SpinNet,
  title={SpinNet: Learning a General Surface Descriptor for 3D Point Cloud Registration},
  author={Ao, Sheng and Hu, Qingyong and Yang, Bo and Markham, Andrew and Guo, Yulan},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2021}
}

References

[1] 3DMatch: Learning Local Geometric Descriptors from RGB-D Reconstructions, Andy Zeng, Shuran Song, Matthias Nießner, Matthew Fisher, Jianxiong Xiao, and Thomas Funkhouser, CVPR 2017.

Updates

03/04/2021: The code is released!
01/03/2021: This paper has been accepted by CVPR 2021!
25/11/2020: Initial release!